Guard screens every prompt, response and tool call in real time: prompt attack, personal data, unsafe content, unknown links and unsafe agent actions. It flags, masks or blocks per your policy. No model or code changes.
Each policy decides the action. PII can be masked in place: the request continues with placeholders, the personal data never reaches the model.
Four levels, L1 to L4, set how strict each detector is. One knob per policy, applied consistently across every detector.
Language is detected per request and routed to per-language engines. Detection quality is measured per language and published in our benchmark.
Repeat verdicts are served from a content-keyed cache in about 10 milliseconds. Under overload Guard queues instead of failing: no request is silently dropped.
Every verdict is written to a tamper-evident audit chain and feeds Comply as evidence. That is the loop: what Guard catches, Comply proves.
The demo is live. Point it at a prompt and watch the verdict.